About your fault, it's time to do some standard troubleshooting: Strip down the system as much as possible. So, all USB devices out except the bare essentials, you can even unplug the front USB cables from the headers. Take out or disconnect unnecessary drives (basically all but the boot drive). Leave only one RAM module in slot A2 or two modules in A2 and B2. Double-check that all cables from the PSU are connected and fully latched on, on both sides (PSU and board/GPU). You could even try sourcing a different PSU and GPU for testing, if that's somehow possible. Basically, detach/unplug as much as possible (except the power cables of course) and see if it helps at all.
Then, you're getting up to 100°C CPU temperatures under load with a capable 360mm AIO, this means you haven't set any reasonable power limits. Selecting "Water cooler" in the BIOS maxes out the power limits, this is not advised with a 12900K, as it can draw around 300W under full multithreaded load, which is very hard to transport away fast enough for any cooler. The CPU surface is very small (even much smaller than the CPU heatspreader) and you can quickly have hotspots which are very difficult to prevent even with nice watercooling. In
this test here, they found the sweet spot of the 12900K to be around 125W, believe it or not. I would start with something around 200W though for the power limits and see how your cooling does.
The power limit is mostly to prevent near-overheat and throttling situations. Most workloads (including gaming) will not utilize all of the cores to the max, so you very rarely have workloads that lose any performance from setting lower power limits, for gaming this is pretty much excluded. So you set the power limits specifically according to your cooling capabilities, i would aim to be in the high 80°Cs for that. The 90°C range would be uncomfortably high for such a powerful AIO, let alone 100°C which is throttling territory.
Here you can read a thread where i helped someone lower the power limits on the 11900K to get it slightly more under control (even though we didn't find the optimal point, because at some point he was already satisfied).
Here is another thread where we do something similar for a 12900K. Although we also stay pretty generous with the power limits. Usually you improve the calculation efficiency (energy spent per calculation job completed) by lowering the power limits a bit more strictly, as in the TechPowerUp article, where the sweet spot was found somewhere around 125W or so.
Quote: "We also looked at power consumption and efficiency at these TDP limits and found that there's A LOT of efficiency to be gained at lower power limits [with the 12900K]. While the default 241/241 configuration is less efficient than all Zen 3 CPUs, the Ryzen 7 5800X is beat as soon as you go below the 200 W limit. At 190/190, the 5600X can no longer keep up. The most energy-efficient configuration turns out to be 75 W, which would make the Core i9-12900K the second most efficient CPU in our test group, only beaten by the Ryzen 9 5950X. Of course, such low limits will drastically reduce performance—you're trading longer runtime for lower overall power usage. The sweet spot is near 125 W, I'd say, but it also depends on the application."
What this shows is that Intel chose the 241W MTP limit mostly to look good in the launch reviews. Even going through the application performance results with a fine toothed comb, you can see that setting the power limits to 190/190W on a 12900K hardly loses any performance, but obviously needs 50W less power than the default 241/241W under very taxing theoretical load.
MSI definitely set the power limits too high, plus they tend to pre-select Water Cooler by default (which means completely maxed out limits, as i said). But even "Tower Cooler" allows way too high power draw for most coolers to handle. Only "Boxed Cooler" should enforce the official Intel power limits. So the "Tower Cooler" is a wasted opportunity to select a good middle ground, like 150-200W. As it is now, it's almost like maxed out limits, just a little less.
We should also check your sensors with
HWinfo64. You can run it and open "Sensors", then expand all sensors by clicking on the little <--> arrows on the bottom. First let it run in idle for a while, so the "minimum" baselines for the values are established. Then produce full CPU load with Cinebench R23. After the CPU temperatures have stabilized at the highest level (let it run for 10 mins), take a screenshot. This will show everything at once. Expand the columns of the sensors a bit so everything can be read.
